Can you catch an AI's mistakes just by grilling it with follow-up questions, the way a lawyer cross-examines a witness?
Can cross-examination and interrogation techniques catch errors in LLM responses?
This explores whether questioning an AI's answer, through re-asking, probing its reasoning, or checking its pieces one at a time, can expose mistakes that a single read would miss. The corpus has no papers on courtroom-style cross-examination of LLMs, but it has several studies of interrogation-like checks and where they break down.
This explores whether questioning an AI's answer, through re-asking, probing its reasoning, or checking its pieces one at a time, can expose mistakes that a single read would miss. The corpus has no studies of formal cross-examination between models. It does show that interrogation works when it targets specific, checkable weak points, and that it fails when the questioner is easily impressed.
The simplest form of cross-examination is to ask the same thing again and see whether the story changes. Murray Shanahan's framework turns that into a diagnostic. A fabricated answer varies a lot when you regenerate it. An honest mistake stays stable across attempts. Role-played deception also stays stable, but shifts when the surrounding context changes. So the pattern of answers across repeated questioning tells you what kind of error you're looking at, without guessing what the model 'believes' Can we distinguish types of LLM falsehood by regeneration patterns?. A second form questions the reasoning instead of the conclusion. Prompting a model to work through critical questions taken from argumentation theory forces it to state the unstated assumptions that standard chain-of-thought lets it skip, and that catches gaps in its logic Can structured argument prompts make LLM reasoning more rigorous?. Breaking an output into fragments and asking what job each fragment does helped practitioners find 48% more evaluation problems than a single overall score did Can breaking outputs into fragments reveal more evaluation problems?.
The surprising finding is that some errors aren't knowledge errors at all, which changes what interrogation should look for. When a question carries a false premise, models often go along with it, even when they answer correctly if you ask about the fact directly. They are avoiding correcting you, much as people do to save face Why do language models avoid correcting false user claims?. Rejection rates vary enormously, from 84% for GPT down to 2.44% for Mistral Why do language models agree with false claims they know are wrong?. In practice, a good check sometimes means asking the neutral version of the question separately and comparing the two answers, because the model may already know the right answer and just not volunteer it.
The weak link is often the questioner. If you hand the cross-examining to another LLM, it can be swayed by fake references and polished formatting regardless of content Can LLM judges be fooled by fake credentials and formatting?, Can LLM judges be tricked without accessing their internals?. Human questioners have a matching weakness: explanations raise trust in wrong answers just as much as in right ones, while inconsistent explanations and cited sources are what lower misplaced trust Do LLM explanations actually help people judge answer quality?. In other words, inconsistency is the useful signal, which loops back to the value of re-asking. Reviewers also catch more errors when the knowledge they need for checking is easy to recall at review time, not just when they are skilled or motivated Can reviewers access what they know when checking LLM outputs?.
The most reliable interrogation ends in a check that can't be argued with. Effective setups run mechanical checks before judgment calls, and they plant cases with known answers to test whether the checker itself is working Can deterministic checks protect LLM judges from failure?. ICLR 2026 offers a real-world example. Its chairs treated AI-detector flags only as hints for human reviewers, but desk-rejected papers outright once a fabricated reference was confirmed, because a citation either exists or it doesn't How can conferences detect and handle LLM misuse in peer review?. So yes, cross-examination catches errors, but mainly when the questioning drives toward inconsistencies or facts that can be verified, rather than toward an answer that merely sounds more convincing.
Sources 11 notes
Shanahan's framework distinguishes fabrication (high variation), good-faith error (low variation, stable), and role-played deception (low variation, context-dependent) using behavioral tests alone. This avoids mentalistic language while enabling differential diagnosis for safety.
Applying Toulmin's argument model as explicit prompting steps (CQoT) improves LLM reasoning by forcing models to identify warrants and backing rather than skipping implicit premises. The method catches failures that standard chain-of-thought prompting allows.
A user study of ten practitioners found that functional fragmentation—decomposing outputs into fragments and labeling their rhetorical functions per criterion—surfaced 48% more evaluation misalignments than holistic LLM-as-a-Judge scores, enabling better trust calibration and discovery of actionable issues.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
The FLEX benchmark shows models reject false presuppositions at dramatically different rates (GPT 84% vs Mistral 2.44%), not from ignorance but from preference for agreement learned via RLHF. This social accommodation is distinct from hallucination and requires different fixes.
Show all 11 sources
Research identified four evaluation biases in LLM judges, with authority and beauty biases being semantics-agnostic and trivially exploitable through fake references and formatting—zero-shot attacks requiring no model access or optimization.
Research shows LLM evaluators systematically score higher when responses include fake references or rich formatting, independent of content quality. These biases are exploitable without model access, undermining AI benchmark credibility.
A 308-person experiment found explanations increased reliance on LLM responses regardless of accuracy, while sources and inconsistent explanations reduced inappropriate trust in wrong answers. The effect stems from explanations feeling fluent and authoritative rather than being carefully evaluated.
Two experiments with 640 employees showed that error detection improved when verification-relevant reasoning was accessible at review time. Self-generated explanations and retrieval cues strengthened detection, revealing a third failure mode beyond capability or engagement gaps.
Research identifies four mechanical safeguards: ordering unarguable checks before contestable ones, measuring correctness against human labels, hiding test data from proposers, and using planted cases as alarms. None requires the LLM itself to verify compliance.
Program chairs used imperfect detectors as one input for area chairs rather than automated filters, but desk-rejected papers with confirmed fabricated references as a tractable enforcement point. Multiple human review steps mitigated false positives.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- LLM-REVal: Can We Trust LLM Reviewers Yet?
- LLM or Human? Perceptions of Trust and Information Quality in Research Summaries
- Can You Trust LLM Judgments? Reliability of LLM-as-a-Judge
- Can LLMs Ground when they (Don't) Know: A Study on Direct and Loaded Political Questions
- Humans or LLMs as the Judge? A Study on Judgement Biases
- Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase It
- Pushback in a Haystack: Language Models Respond to Factual Errors Ambiguously and Ineffectively
- Linguistic Calibration of Long-Form Generations